Query Rewriting for Semantic Web Information Integration Dave Kolas

advertisement
Query Rewriting for Semantic Web Information Integration
Dave Kolas
BBN Technologies
1300 N 17th St., Suite 400, Arlington, VA, U.S.A.
dkolas@bbn.com
This richness of data definition offers significant benefits
for information integration. Queries can be expressed in a
more abstract fashion, and more domain knowledge can be
encapsulated in the data definition. This means that the
information consumer spends less time determining the
appropriate vocabulary for their query and thus has more
time to take action on the results.
The goal of this paper is to apply existing work on information integration systems to an information integration
system using Semantic Web technologies. This system
will use OWL ontologies instead of schema definitions and
SWRL rules instead of view definitions. We will examine
exactly what is desired from such a system, and then explore how current approaches can be extended to achieve
those goals.
The rest of the paper will be structured as follows.
First, we will consider the larger picture of an architecture
in which this type of semantic query rewriting would be
necessary. Next, we will attempt to apply standard techniques for schema mediation, Global as View and Local as
View, to this scenario, analyzing the strengths and weaknesses of each approach. Finally, we will propose modifications of the Global as View approach which overcomes
the weaknesses, and analyze their effects on query rewriting.
Abstract
Though significant research has been done in the realm
of information integration, continuously increasing
amounts and complexity of data demand more advanced
techniques for its access. Semantic Web technologies
such as OWL and SWRL provide the capacity for rich
definitions of data sources, global views, and the mappings between them. An information integration system
based on these technologies is thus very attractive;
however, OWL and SWRL would appear to introduce
significant complexity to query reformulation. We motivate our work via a distributed query system architecture
based on Semantic Web technologies, and discuss the
application of Global as View and Local as View approaches to this scenario. Finally, we introduce modifications to the Global as View approach for this system and
analyze their effects.
Introduction
As the amount of data available on the Web and from other
sources explodes, there is an ever increasing need for more
advanced data integration techniques. Organizations that
assimilate this information for processing are left drowning
in the sheer volume of this data. An organization’s perspective on available data is constantly evolving, and as
such their information integration systems must also
evolve. New sources of data appear constantly, and existing sources change or disappear. If the information integration systems cannot quickly adapt to these types of
changes, their value becomes negligible.
Significant advances in the realm of information retrieval
have made searching the text of documents significantly
more effective. However, these techniques are primarily focused on unstructured documents and documents whose
structure are designed for human visualization, and cannot
hope to do more than point the consumer to relevant
documents. As such, they are not well suited for answering structured queries over structured sources.
The goal of the Semantic Web is to bridge this gap by
making information on the Web interpretable by computers. Common languages such as RDF [RDF], OWL
[OWL], and SWRL [SWRL] were created to allow rich
data definition in a standard language that could be processed in a well-defined way. OWL, based in Description
Logic, and SWRL, based in Description Logic and Logic
Programming, when used together provide a basis for far
more expressive data descriptions than are currently possible with web data formats such as XML.
Semantic Distributed Query Architecture
The semantic distributed query architecture is motivated
by a desire to provide data integration through flexible
mappings, created in the form of ontologies and rules. The
system makes use of both Description Logic-based descriptions in OWL and Horn-like rules in SWRL. This leads
to two specific aims for the system design. First, make
use of the expressivity of OWL and SWRL to allow for a
level of conceptual data independence beyond that which
can be provided by non-semantic descriptions. Second,
the mapping between the data sources and the semantic
system should be done in such a way that the underlying
data storage mechanisms are completely irrelevant to the
consumer.
To facilitate these goals, the architecture splits the query
and data transformations into two distinct tiers of processing (See Figure 1). The upper tier of the architecture is responsible for ontology to ontology transformations,
subquery ordering, and information integration, commonly
referred to as the mediator. The component that fulfills
this role is referred to as Semantic Query Decomposition
44
(SQD). The lower tier is responsible for encapsulating native data sources as ontological knowledge with a
SPARQL [SPARQL] interface.
This tier must map
SPARQL queries into queries appropriate for a particular
native data source (e.g. SQL for relational databases). This
tier is manifested in Semantic Bridges for each type of data
source, and corresponds with the well known concept of a
wrapper [Ullman97]. We will now look at SQD and two
instances of Semantic Bridges in more detail.
Se mantic Bridge for Web Services
Another instance of the Semantic Bridge is the Semantic
Bridge for Web Services (SBWS). The purpose of this
component is to serve as a wrapper over a web service with
an interface defined by WSDL. Unlike SBRD, SBWS is
not mapping one general query language to another, but
rather mapping one general query language onto a set of
specific predefined operations. This has proven challenging, though precedent does exist [LRO96]. Again, we will
assume that such a thing is possible and that the only relevant problem for the mediator is producing an appropriate
SPARQL query.
Applying Mediated Schema Approaches
Semantic Query Decomposition (SQD)
Semantic Bridge
Relational Databases
RDBMS
Semantic Bridge
Web Services
Data
Source
Two primary approaches [Levy00] exist for restricting the
mappings between local schemas and the target schema in
order to make the query rewriting process more tractable.
These are Global-as-View (GAV), where elements in the
target schema is defined as a view over the local schemas,
and Local-as-View (LAV), where elements in the local
schemas are defined as a view over the global schema. We
will examine each of these in turn, analyzing how each applies to our semantic distributed query architecture, and the
practical downsides to each. First, however, it is useful to
examine the common elements of applying these approaches to Semantic Web information integration system.
Because SWRL as an addition to OWL is essentially
datalog restricted to unary and binary predicates, much of
the literature on information integration approaches applies
to it quite directly. Some of the vocabulary differs,
though: whereas datalog descriptions generally refer to
predicates, SWRL rules, with their heritage in Horn rules,
generally refer to atoms. Though SWRL is technically a
superset of OWL, most of the translation rules that have
been created in practice make use of this datalog portion of
the language. As such, that will be the focus of discussion
in this paper, excepting the section on future work.
Thus, these SWRL rules directly correspond to the view
definitions found in the literature on both the GAV and
LAV approaches. One exception to this correspondence is
that SWRL rules are permitted to have multiple atoms in
the head (consequent):
Semantic Bridge
Relational Databases
RDBMS
Figure 1
Se mantic Query Decomposition
The goal of the upper tier of the architecture is that of the
mediator in standard information integration systems.
This involves receiving the query, deriving a query plan
over the various sources, and optimizing and executing
this plan. Unlike a relational system, however, the set of
inputs to this process are based in Semantic Web technologies. A query Q is defined in the language SPARQL,
using terms from the domain ontology, O, defined in
OWL. Data source descriptions are defined as data source
ontologies, D1 ...Dn , also defined in OWL. Mappings from
the data source ontologies into the domain ontology
M1..Mn are expressed as sets of rules defined in SWRL.
This mediator then must derive a set of subqueries
Qx1 …Qxm where x1..xm ∈{D1...Dn} in SPARQL whose combined results are equivalent to those of Q.
Pie(x) ∧ containsFruit(x, y) ⇒ Dessert(x) ∧ containsIngredient(x,y)
Se mantic Bridge for Relational Databases
One instance of a Semantic Bridge is the Semantic Bridge
for Relational Databases (SBRD). This component is designed to connect to a relational database and present its
contents according to a data source ontology. This implies defining a mapping between the data source ontology
and the underlying relational structure. Work has been
done by the Semantic Web community in mapping
SPARQL queries over a virtual RDF graph to SQL queries
over a relational database given an RDF to RDBMS mapping [D2RQ], and the process for doing such is beyond the
scope of this paper.
However, this rule structure does not add any complexity
to the processing, as these rules can simply be rewritten as
multiple rules with single-atom consequents:
Pie(x) ∧ containsFruit(x, y) ⇒ Dessert(x)
Pie(x) ∧ containsFruit(x,y) ⇒ containsIngredient(x,y)
Now we will examine how the GAV and LAV approaches apply to a Semantic Web system individually.
45
tates practical usability in a real-world scenario. We will
consider several of these and discuss why each is desirable.
First, the mediated schema should be defined independently. There are many reasons why this is desirable. If the
mediated schema is built independently, it is much less
likely to reflect the biases of any of the individual sources
relative to one another. It is also more likely to be free of
the limitations of whatever underlying access mechanism is
being used for each of the sources. Moreover, the users’
perspective is likely to change over time. Keeping the
mediated schema independent allows change of this perspective while minimally affecting the data source mappings.
Second, mappings from each source to the mediated
schema should be mutually independent. This is critical,
as any information integration system is only going to
grow in complexity over time. As more useful sources of
data are discovered, the complexity of adding them cannot
increase or the approach will not be scalable. Over time, it
is also likely that sources will be replaced with updated
versions of themselves. If removing, adding, or replacing a
data source requires revisiting the mappings between every
data source and the target, maintenance of the system
would become unwieldy very quickly.
Third, the mappings should be able to leverage all of the
meaning stored in the sources. There tends to be a mismatch between the Semantic Web technologies, which are
primarily based on monotonic reasoning and an open
world assumption, and typical relational database implementations, in which the absence of values can often be interpreted as having meaning. While this is not necessarily
a general information integration problem, it is a problem
that the system will have to overcome in order to be effective.
Finally, the mappings should be as straightforward to
create as possible. Ideally, these mappings could be created by a domain expert and not require a software engineer. While this is notably more difficult to quantify than
the previous goals, and undoubtedly a matter of opinion, it
is our belief that writing source to target transformations is
easier than defining sources in terms of the global schema.
Global as View
Applying the GAV approach to our system based on OWL
ontologies and SWRL rules means treating the SWRL
rules as view definitions as above. In the GAV approach,
each mediated schema predicate is implied by a conjunction of data source predicates. This means that for every
predicate in the domain ontology, a direct derivation is required directly from data source concepts.
The GAV approach has been shown to have both positive and negative aspects with respect to information integration systems [Levy00], and we believe these apply
equally to a Semantic Web based information integration
system. On the positive side, query rewriting is straightforward, and since the portion of SWRL used for the mappings is a subset of datalog, this is true in this scenario as
well. Essentially all that needs to be done is replacing
mediated schema predicates in the query with the data
source predicates that imply them. Also, GAV statements
follow a natural mapping paradigm; they follow the format
of if some set of predicates are true in the data source, then
some set of predicates are true in the mediated schema.
On the other hand, GAV has some significant disadvantages. First, the mediated schema is dependent upon the
data source schemas for structure. Also, mappings from
the data sources to the mediated schema are not mutually
independent. This is generally thought to limit the scalability of the GAV approach, and it is equally limiting
when using Semantic Web technologies.
Local as View
In the LAV approach, the mapping descriptions are reversed. The contents of the data sources are described in
terms of predicates on the mediated schema. Once again,
we could use SWRL rules to represent these relationships.
A LAV approach has been proven successful in a Semantic
Web context, though it made use of only OWL predicates
for alignment and not SWRL rules [DHQW06].
The LAV approach has two major advantages over the
GAV approach. First, the mediated schema is more independent. Because local sources are defined in terms of the
mediated schema, it is possible to start with whatever mediated schema is desired. Perhaps more importantly, the
data source mappings are mutually independent. There is
no interaction between data sources in the definitions of the
source mappings, and thus a new source can be added,
modified, or removed without affecting any of the other
mappings.
Unfortunately, this approach makes query rewriting significantly more complex. Algorithms for this rewriting
have been developed [Levy00, PH01], but even then recursive queries may be required [Levy00].
Modified GAV Approach
With the desire for an independent global schema and
independent source mappings, it may seem as though we
are trending towards an approach based on LAV. However, we feel that the desire to write the mappings as
source-to-target is extremely important, and perhaps more
importantly that the other considerations can be accommodated within a modified GAV approach.
It is worth noting that it has been shown that under certain circumstances one need not choose between LAV and
GAV approach. In the GLAV approach, mappings can be
defined in either direction, provided the global schema
meets certain conditions [FLM99, CCGL02]. Unfortunately, the primary condition that the global schema has to
meet for GAV and LAV to be interchangeable is the inclu-
Information Integration Goals
There are several criteria that define a successful approach
to the data integration problem in this context. Each facili-
46
sion of integrity constraints. Since OWL and SWRL have
no capacity for integrity constraints, this requirement could
not be met without fundamentally altering the spirit of
OWL/SWRL based reasoning. Other approaches exist for
combining them as well, but are based on relational algebra and thus less directly applicable to this scenario [XE].
With these things considered, we now present an application of modified GAV to a Semantic Web information
integration system. For the purposes of illustration, we
will use a simple example, with three data sources, and a
domain ontology:
Mapping2 states that if the delivery service has delivered a
pie from Dave’s Desserts that contains a fruit, the store
stocks a pie with that fruit. Mapping3 states that if a customer has ordered a pie that contains a fruit, the customer
must like that fruit (the pie aficionados are unyielding in
their belief that no person orders a pie with a fruit that they
do not at least subconsciously like). This type of rule is
not necessarily a problem with only three sources, but
when more similar sources are added, the complexity explodes. Adding another pie supplier with the same schema
means doubling the number of mappings like Mapping3.
Adding another pie supplier, another dessert shop, and another delivery service means Mapping3 has now become
eight different mappings. The maintenance cost increases
further if the schemas for these new sources are subtly (or
vastly) different.
These complications are derived from the restriction
noted before that all antecedent terms in the GAV approach
must be derived only from data source predicates. By lifting this restriction and imposing two new restrictions, we
can allow the data source mappings to be defined independently.
Pie Supplier: (pie:) “Pete’s Pies” is a commercial
suppler of pies
Dessert Shop: (shop:) “Dave’s Desserts” is a retail
outlet selling pies and other items, and purchases
their pies from Pete’s.
Delivery Service: (del:) “Don’s Delivery” is a delivery service that works with Dave’s Desserts as well as
other retail stores
Domain Ontology: (pa:) represents the mediated
schema for the three sources, assumed to be developed
by an organization of pie aficionados
[R1] All predicates appearing in the antecedent of
mapping m ∈ Md are in O ∪ Dd
[R2] All predicates appearing in the consequent of
mapping m ∈ Md are in O
A full description of the predicates in these three ontologies
is in the Appendix. Note that this scenario assumes an
ideal world in which pie can be delivered to one’s door.
This implies that while each mapping’s consequent must
still be expressed in the predicates of the domain ontology,
its antecedent may now be comprised of predicates both
from the associated single data source and the domain ontology. However, predicates from multiple data sources are
no longer allowed to appear in the same mapping.
Queries could only be answered, however, if these mappings are acyclic:
Independent Data Sources
Perhaps the most glaring problem with applying GAV to
an information integration system is non-independent
source to target mappings. When only one data source is
required for a given predicate in the domain ontology, this
is not a problem. Consider the mapping:
[Mapping1]
pie:Pie(x)
pie:containsFruit(x,”apple”) ⇒ pa:ApplePie(x)
∧
[R3] For each predicate used in the mappings, assign
a node in a graph. For each pair of predicates such that
one appears in the antecedent and one appears in the
consequent of the same mapping, assign a directed
link from the antecedent node to the consequent node.
The resulting graph must be acyclic.
This mapping is likely to be unaffected by the addition,
deletion, or modification of other data sources. However,
rules that come from multiple sources such as these may
not be so easy:
Since this condition can be checked for at design time,
queries stuck in recursive loops can be prevented. It is
easy to see then that mappings of this type can be “unfolded” into view definitions in the standard GAV approach.
The implications to the maintenance of the system,
however, are more noteworthy. Consider a possible new
formulation of Mapping3:
[Mapping2] pie:Pie(p) ∧ pie:containsFruit(p,f) ∧
pie:name(n) ∧ shop:Dessert(d) ∧ shop:name(d,n) ∧
∧
del:productOrdered(o,p)
∧
del:Order(o)
∧
del:Shop(s)
∧
del:orderedFrom(o,s)
del:shopName(s,”Dave’s Desserts”) ∧ del:Product(p)
del:productName(p,n) ⇒ pa:stocksPieWithFruit(s,f)
[Mapping3] pie:Pie(p) ∧ pie:containsFruit(p,f) ∧
pie:name(n) ∧ shop:Dessert(d) ∧ shop:name(d,n) ∧
del:Order(o) ∧ del:orderedBy(o,c) ∧ del:Customer(c)
∧ del:productOrdered(o,p) ∧ del:orderedFrom(o,s) ∧
del:Shop(s) ∧ del:shopName(s,”Dave’s Desserts”) ∧
∧
del:productName(p,n)
⇒
del:Product(p)
pa:likes(c,f)
[Mapping3.1] pie:Pie(p) ∧
pa:Pie(p) ∧ pa:pieName(p,n)
pie:name(p,n)
⇒
[Mapping3.2] del:Shop(s) ∧ del:shopName(s,n) ⇒
pa:Shop(s) ∧ pa:shopName(s,n)
47
[Mapping3.3] pa:Shop(s) ∧ pa:shopName(s,”Dave’s
∧
shop:Dessert(d)
∧
pa:Pie(p)
Desserts”)
∧
shop:name(d,n)
⇒
pa:pieName(p,n)
pa:sellsPie(s,p)
[1] A predicate p may be negated in the antecedent of
a mapping m ∈ Md if and only if p ∈ Dd
[2] No predicates may be negated in the consequent of
a mapping.
[Mapping3.4] del:Order(o) ∧ del:orderedBy(o,c) ∧
del:Customer(c) ∧ del:productOrdered(o,pr) ∧
∧
del:productName(pr,n)
∧
del:Product(pr)
del:orderedFrom(o,s) ∧ pa:Shop(s) ∧ pa:sellsPie(s,p)
∧
pa:Pie(p)
∧
pa:pieName(p,n)
⇒
pa:orderedPie(c,p)
This is a reasonable extension because these predicates can
only be leaf nodes in the unfolding of the original query,
and SPARQL is capable of expressing negation through a
combination of an “OPTIONAL” clause and the use of the
filter “bound”. More formally, for any predicate p that is
in need of negation, we can define a predicate p’ to mean
the absence of p. Since SPARQL allows querying for negation, when the system would query for p’, the system can
replace it with the SPARQL construction for ℜp. Thus
this restriction of negation does not add any complexity to
the query rewriting process. However, this does imply
that our mappings are outside of SWRL as it is currently
defined. Since it is yet to be seen whether and in what
form negation will be supported in whatever Semantic
Web rule language is created by the W3C Rule Interchange
Format Working Group [RIF], it seems reasonable to incorporate it conceptually into an information integration
system.
The result is negation of predicates in the underlying
sources is supported, but the monotonicity of the domain
ontology is preserved.
[Mapping3.5] pa:Pie(p) ∧ pa:orderedPie(c,p) ∧
pie:containsFruit(p,f) ⇒ pa:likes(c,f)
This is only one such possible formulation. These restrictions on the mappings result in new, different mappings
that are naturally more adaptable to changes in the system.
For instance, if Pete’s Pies changes their schema, only
M3.1 and M3.5 require changes, regardless of how many
retail shops or delivery services exist.
One minor disadvantage of this technique is that some
intermediate predicates that may otherwise not have been
needed in the domain ontology are required. For instance,
if the pie aficionados were only concerned with which people liked which fruits, and not with which shops sold
which pies, the predicate pa:sellsPie would be unnecessary
from the perspective of the domain ontology design, but
necessary for the mapping sets to be successfully divided.
In practice, however, this type of intermediate predicate
could be filtered from the results and disallowed in queries,
resulting in it effectively not being in the visible domain
ontology.
Conclusions
Based on our work, it is clear that previous research on information integration approaches is entirely applicable to
an information integration system based on Semantic Web
technologies when data source to domain ontology mappings are represented in SWRL. This is because SWRL
is a subset of datalog, which is referenced in much of the
information integration literature.
Further, we have shown that we can use a modification
of the GAV approach to model our information integration
system in a way that preserves the stated goals: an independent domain ontology, mutually independent data
source to domain ontology mappings, retention of all
meaning from data source to domain ontology, and a natural mapping between data source and domain ontologies.
We accomplished these goals by starting with the GAV
approach and disallowing predicates from multiple data
sources in the mappings’ antecedents and instead allowing
both data source and domain ontology concepts to appear
there. This provided both mutually independent data
source mappings and a more independent domain ontology. Finally, we added a limited form of negation to the
mapping rules, allowing the retention of meaning based on
the absence of statements in the data sources. The result is
a sound method of utilizing SWRL in place of relational
view definitions in support of a Semantic Web based information integration system.
Independent Domain Ontology
In order to have a system in which the users’ perspective
can evolve over time, it is necessary to have a domain ontology whose definition is independent of that of the data
sources. Fortunately, this is enabled almost as a side effect
of the independence of the data source mappings outlined
above. If one defines the domain ontology before any of
the mapping from the data sources is started, and each data
source maps into that ontology independently of the others, then a change to the domain ontology is guaranteed to
affect the smallest possible number of mappings.
Negation
Finally, we address a problem that is unique to systems
based on Semantic Web technologies. As noted before,
the majority of these technologies are based on the open
world assumption and monotonic reasoning. As such,
there is no place for deriving conclusions from the absence
of statements. However, other types of sources that one
might typically want to integrate operate on a closed world
assumption, and thus the lack of a statement may in fact be
meaningful.
We address this issue by allowing a restricted form of
negation in the rules.
48
th
Semantic Web System. In Proceedings of the 5 International Semantic Web Conference. Springer.
Future Work
Our continued work in this area is divided along two distinct paths. First, we hope to expand the approach and algorithm to allow for more abstract concept definitions.
Second, we have built and are enhancing a prototype system based on the design discussed above.
We hope to extend the approach to make better use of
the OWL DL constructs in the domain and data source ontologies. Since many OWL DL reasoners are based on reducing equivalent OWL axioms into LP rules, it seems
appropriate to extend the query decomposition reasoning in
this direction. Naturally, not all OWL axioms can be expressed in LP [GHVD03], but utilizing those that are is a
reasonable starting point. By adding these rules to the
query rewriting process, we could begin querying the information integration system with DL constructs, e.g. class
intersection and disjunction, etc.
We are also currently working on implementing the
three pieces of software SQD, SBRD, and SBWS. This
should lead to a proof of concept system which will prove
the validity of the approach in the context of a real world
example.
Klyne, G., and Carroll, J., eds. 2004. Resource Description Framework (RDF): Concepts and Abstract Syntax.
http://www.w3.org/TR/rdf-concepts/
Dean, M., and Schreiber, G., eds. 2004. OWL Web Ontology Language Reference.
http://www.w3.org/TR/owl-ref/
Horrocks, I., Patel-Schneider, P., Boley, H., Tabet, S.,
Grosof, B., and Dean, M. 2004. SWRL: A Semantic
Web Rule Language Combining OWL and RuleML.
http://www.w3.org/Submission/SWRL/
W3C Rule Interchange Format
http://www.w3.org/2005/rules/
Working
Group.
Prud’hommeaux, E., Seaborne, A. 2006. SPARQL Query
Language for RDF.
http://www.w3.org/TR/rdf-sparql-query/
Grosof, B., Horrocks, I., Volz, R., and Decker, S. 2003.
Combining Logic Programs with Description Logic. In
th
Proceedings of 12 International Conference on the
World Wide Web. New York, NY: ACM Press.
References
Calì, A., Calvanese, D., De Giacomo, G., and Lenzenrini,
M. 2002. On the Expressive Power of Data Integration
Systems. London, UK: Springer-Verlag.
Appendix: Example System Definitions
Levy, A. 2000. Logic-Based Techniques in Data Integration. Norwell, MA: Kluwer Academic Publishers.
Following are the definitions of the data sources and domain ontologies referenced above, showing classes, properties and cardinalities.
Ullman, J. 1997. Information Integration using Logical
Views. London, UK: Springer.
Pie Supplier: (pie:)
• Pie(name, containsFruit*)
Dessert Shop: (shop:)
• Dessert(name)
Delivery Service: (del:)
• Shop(shopName)
• Order(orderedBy, productOrdered*,orderedFrom)
• Customer()
• Product(productName)
Domain Ontology: (pa:)
• Pie(pieName)
• ApplePie(pieName)
• Shop(shopName, sellsPie*, stocksPieWithFruit*)
• Person(orderedPie*,likes*)
Levy, A., Rajaraman, A., and Ordille, J. 1996. Querying
Heterogeneous Information Sources Using Source Descriptions. In Proceedings of the Twenty-second International
Conference on Very Large Databases. Bombay, India:
VLDB Endowment, Saratoga, California.
Friedman, M., Levy, A., and Millstein, T., 1999. Navigath
tional Plans for data integration. In Proc. Of the 16
Nat. Conf. on Artificial Intelligence, p 67-73. AAAI
Press/The MIT Press.
Xu, L., and Embley, D. Combining the Best of Global-asView and Local-as-View for Data Integration. citeseer.ist.psu.edu/604666.html
Bizer, C., and Seaborne, A., 2004. D2RQ – Treating
Non-RDF Databases as Virtual RDF Graphs. In Proceedrd
ings of the 3 International Semantic Web Conference.Springer.
Dimitrov, D., Heflin, J., Qasem, A., and Wang, N. 2006.
Information Integration Via an End-to-End Distributed
49
Download